Embed Size (px)
Citation preview
1
Spatial Big Data (SBD) Challenges Intersecting Mobile and Cloud Computing
Shashi Shekhar McKnight Distinguished University Professor
Department of Computer Science and Engineering
University of Minnesota
www.cs.umn.edu/~shekhar
2
Spatial Big Data (SBD)
• SBD are important to society
– Ex. Eco-routing, Public Safety & Security, Understanding Climate Change
• SBD exceed capacity of current computing systems
• DBMS Challenges
– Eco-Routing: Lagrangian frame, Non-Stationary Ranking
– Privacy vs. Utility Trade-offs
• Data Analytics Opportunities
– Post Markov Assumption – Estimate Neighbor Relationship from SBD
– Place based Ensemble Models to address spatial heterogeneity
– Bigger the spatial data, simpler may be the spatial models
– Online Spatial Data Analytics
• Platform Challenges
– Map-reduce – expensive reduce not suitable for iterative computations
– Load balancing is harder for maps with polygons and line-strings
– Spatial Hadoop ?
3
Relational to Spatial DBMS to SBD Management
• 1980s: Relational DBMS • Relational Algebra, B+Tree index
• Query Processing, e.g. sort-merge equi-join algorithms, …
• Spatial customer (e.g. NASA, USPS) faced challenges
• Semantic Gap
• Verbose description for distance, direction, overlap
• Shortest path is Transitive closure
• Performance challenge due to linearity assumption
• Are Sorting & B+ tree appropriate for geographic data?
• New ideas emerged in 1990s • Spatial data types and operations (e.g. OGIS Simple Features)
• R-tree, Spatial-Join-Index, space partitioning, …
• SBD may require new thinking for
• Temporally detailed roadmaps
• Eco-routing queries
• Privacy vs. Utility Trade-off
4
Data Mining to Spatial Data Mining to SBD Analytics
• 1990s: Data Mining
• Scale up traditional models (e.g., Regression) to large relational databases (460 Tbytes)
• New pattern families: Associations : Which items are bought together? (Ex. Diaper, beer)
• Spatial customers
• Walmart: Which items are bought just before/after events, e.g. hurricanes?
• Where is a pattern (e.g., (diaper-beer) prevalent?
• Global climate change: tele-connections
• But faced challenges
• Independent Identical Distribution assumption not reasonable for spatial data
• Transactions, i.e. disjoint partitioning of data, not natural for continuous space
• This led to Spatial Data Mining (last decade)
• SBD raise new questions • May SBD address open questions, e.g. estimate spatial neighborhood (e.g., W matrix)?
• Does SBD facilitate better spatial models, e.g., place based ensembles beyond GWR?
• (When) Does bigger spatial data lead to simpler models, e.g. database as a model ?
• On-line Spatio-temporal Data Analytics
5
Parallelizing Spatial Big Data on Cloud Computing
• Case 1: Compute Spatial-Autocorrelation Simpler to Parallelize
– Map-reduce is okay
– Should it provide spatial de-clustering services?
– Can query-compiler generate map-reduce parallel code?
• Case 2: Harder : Parallelize Range Query on Polygon Maps
– Need dynamic load balancing beyond map-reduce
– MPI or OpenMP is better!
• Case 3: Estimate Spatial Auto-Regression Parameters, Routing
– Map-reduce is inefficient for iterative computations due to expensive “reduce”!
– MPI, OpenMP, Pregel or Spatial Hadoop is essential!
– Ex. Golden section search, Determinant of large matrix
– Ex. Eco-routing algorithms, Evacuation route planning
8
8
Smarter
Planet
SIG
SPATIAL
Spatial Computing: Reccent Trends
9
Motivation for Spatial Big Data (SBD)
• Societal: • Google Earth, Google Maps, Navigation, location-based service
• Global Challenges facing humanity – many are geo-spatial!
• Many may benefit from Big Spatial Data
10
Emerging SBD: Geo-social Media, Device2Device
11
Outline
• Motivation
• What is Spatial Big Data (SBD)? • Definitions
• Examples & Use Cases
• SBD Infrastructure
• SBD Analytics
• Conclusions
12
Spatial Big Data Definitions
• Spatial datasets exceeding capacity of current computing systems • To manage, process, or analyze the data with reasonable effort
• Due to Volume, Velocity, Variety, …
• SBD History • Data-intensive Computing: Cloud Computing, Map-Reduce, Pregel
• Middleware
• Big-Data including data mining, machine learning, …
•SBD may challenge • DBMS
• Conceptual Data Model, e.g., Frames of Reference
• Query Processing, e.g., Non-Stationary Ranking of candidates
• Data Analytics
• Lack of transactions
• Lack of independent samples and identical distributions
• Cloud Platform, e.g., map-reduce • Expensive reduce not suitable for iterative computation
• Load balancing is harder for polygons and line-strings
13
Traditional Spatial Data
• Spatial attribute:
– Neighborhood and extent
– Geo-Reference: longitude, latitude, elevation
• Spatial data genre
– Raster: geo-images e.g., Google Earth
– Vector: point, line, polygons
– Graph, e.g., roadmap: node, edge, path
Raster Data for UMN Campus
Courtesy: UMN
Vector Data for UMN Campus
Courtesy: MapQuest
Graph Data for UMN Campus
Courtesy: Bing
14
Raster SBD
LiDAR & Urban Terrain
Change Detection Feature Extraction
Average Monthly Temperature
(Courtsey: Prof. V. Kumar)
• Data Sets >> Google Earth
– Geo-videos from UAVs, security cameras
– Satellite Imagery (periodic scan), LiDAR, …
– Climate simulation outputs for next century
• Example use cases
– Change detection, Feature extraction,Urban terrain
15
Vector SBD from Geo-Social Media
• Vector data sub-genre
– Point: location of a tweet, Ushahidi report, checkin, …
– Line-strings, Polygons: roads in openStreetMap
• Use cases: Persistent Surveillance
– Outbreaks of disease, Disaster, Unrest, Crime, …
– Hot-spots, emerging hot-spots
– Spatial Correlations: co-location, teleconnection
16
Persistent Surveillance at American Red Cross
• Even before cable news outlets began reporting the tornadoes that ripped through Texas on
Tuesday, a map of the state began blinking red on a screen in the Red Cross' new social
media monitoring center, alerting weather watchers that something was happening in the
hard-hit area. (AP, April 16th, 2012)
17
Emerging SBD: Geo-Sensor Networks
17
• Geo-Sensor Network Examples
– Urban roads: Loop detetors, Cameras
– Electricity distribution grids, …
– Environmental sensors for air quality, …
– Robot with sensors, …
• Sensor Network sub-genre
– Fixed reasonable resource: traffic sensors
– Ad-hoc, resource poor: wireless sensor networks
• Use cases:
– Monitor events, anomalies, e.g., accidents,
Congestion, hotspots
– Feed-back control, Predictive planning
– Environmental Health
18
Graphs SBDs: Temporally Detailed
• Spatial Graphs, e.g., Roadmaps, Electric grid, Supply Chains, …
– Temporally detailed roadmaps [Navteq]
• Use cases: Best start time, Best route at different start-times
19
Emerging SBD: Mobile Device2Device
19
• Mobile Device
– Cell-phones, cars, trucks, airplanes, …
– RFID-tags, bar-codes, GPS-collars, …
• Trajectory & Measurements sub-genre
– Receiver: GPS tracks, …
– System: Cameras, RFID readers, …
• Use cases:
– Tracking, Tracing,
• Improve service, deter theft …
– Geo-fencing, Identify nearby friends
– Eco-routing
20
Emergin Use-Case: Eco-Routing
U.P.S. Embraces High-Tech Delivery Methods (July 12, 2007)
By “The research at U.P.S. is paying off. ……..— saving roughly three million
gallons of fuel in good part by mapping routes that minimize left turns.”
• Minimize fuel consumption and GPG emission
– rather than proxies, e.g. distance, travel-time
– avoid congestion, idling at red-lights, turns and elevation changes, etc.
21
Eco-Routing Questions
• What are expected fuel saving from use of GPS devices with static roadmaps?
• What is the value-added by historical traffic and congestion information?
• How much additional value is added by real-time traffic information?
• What are the impacts of following on fuel savings and green house emissions?
– traffic management systems (e.g. traffic light timing policies),
– vehicles (e.g. weight, engine size, energy-source),
– driver behavior (e.g. gentle acceleration/braking), environment (e.g. weather)
• What is computational structure of the Eco-Routing problem?
Predictable
Future
Unpredictable
Future
Stationary ranking
of alternate paths
Non-stationary
ranking of paths
Dijkstra’s, A*, Dynamic Programming….
New Computer Science ?
22
Time Dependence of Shortest Path
22
INPUT:
Source: University of Minnesota
Destination: MSP Airport
Time Interval 7:00am ---12:00noon
OUTPUT:
23
Routing Algorithm Challenges
Non Stationarity ranking
of paths Non FIFO Behavior
Violation of stationary
assumption dynamic
programming
*Flight schedule between Minneapolis and
Austin (TX)
Violates the no wait
assumption of Dijkstra/A*
24
Routing Challenges: Lagrangian Frame of Reference
Question: What is the cost of Path <A,C,D> with start-time t=2 ?
Is it 4 or 5 ? Path T = 0 T = 1 T = 2 T = 3
<A,C,D> 4 3 5 4
<A,B,D> 6 5 4 3
Lagrangian Graph
Snapshots of a Graph
25
Challenges: New Routing Questions
Example Time-Variant Network Questions
New Routing Questions
Best start time to minimize time spend on network
Account for delays at signals, rush hour, etc.
26
• Weekday GPS track for 3 months
– Patterns of life
– Usual places and visits
– Rare places, Rare visits
Work
Home Club
Farm
Morning 7am – 12am
Afternoon 12noon – 5pm
Evening 5pm – 12pm
Midnight 12midnight –
7pm
Total
Home 10 2 15 29 54
Work 19 20 10 1 50
Club 4 5 4 15
Farm 1 1
Total 30 30 30 30 120
SBD Challenge: Privacy vs. Utility Trade-off
27
• Checkin Risks: Stalking, GeoSlavery, …
• Ex: Girls Around me App (3/2012), Lacy Peterson [2008]
• Others know that you are not home!
The Girls of Girls Around Me. It's doubtful any of
these girls even know they are being tracked. Their names and locations have been obscured for privacy
reasons. (Source: Cult of Mac, March 30, 2012)
SBD Challenge: Privacy vs. Utility Trade-off
28
Relational to Spatial DBMS to SBD Management
• 1980s: Relational DBMS • Relational Algebra, B+Tree index
• Query Processing, e.g. sort-merge equi-join algorithms, …
• Spatial customer (e.g. NASA, USPS) faced challenges
• Semantic Gap
• Verbose description for distance, direction, overlap
• Shortest path is Transitive closure
• Performance challenge due to linearity assumption
• Are Sorting & B+ tree appropriate for geographic data?
• New ideas emerged in 1990s • Spatial data types and operations (e.g. OGIS Simple Features)
• R-tree, Spatial-Join-Index, space partitioning, …
• SBD may require new thinking for
• Temporally detailed roadmaps
• Eco-routing queries
• Privacy vs. Utility Trade-off
29
Outline
• Motivation
• SBD Definition and Examples
• SBD Infrastructure • Parallelizing Spatial Computations
• Implications for Cloud Platforms
• SBD Analytics
• Conclusions
30
Parallelizing Spatial Big Data on Cloud Computing
• Case 1: Compute Spatial-Autocorrelation Simpler to Parallelize
– Map-reduce is okay
– Should it provide spatial de-clustering services?
– Can query-compiler generate map-reduce parallel code?
• Case 2: Harder : Parallelize Range Query on Polygon Maps
– Need dynamic load balancing beyond map-reduce
– MPI or OpenMP is better!
• Case 3: Estimate Spatial Auto-Regression Parameters, Routing
– Map-reduce is inefficient for iterative computations due to expensive “reduce”!
– MPI, OpenMP, Pregel or Spatial Hadoop is essential!
– Ex. Golden section search, Determinant of large matrix
– Ex. Eco-routing algorithms, Evacuation route planning
31
Simpler: Land-cover Classification
• Multiscale Multigranular Image Classification into land-cover categories
Inputs Output at 2 Scales
)(2)|()(
)},({maxargˆ
MpenaltyMnobservatiolikelihoodMquality
whereModelqualityodelMModel
32
Parallelization Choice
1. Initialize parameters and memory
2. for each Spatial Scale
3. for each Quad
4. for each Class
5. Calculate Quality Measure
6 end for Class
7. end for Quad
8. end for Spatial Scale
9. Post-processing
0
0.25
0.5
0.75
1
1 2 4 8
Number of Processors
Eff
icie
nc
y
Class-level
Quad-level
0
1
2
3
4
5
6
7
2 4 8Number of Processors
Sp
ee
du
p
Class-level
Quad-level
Input • 64 x 64 image (Plymouth County, MA)
• 4 classes (All, Woodland, Vegetated,
Suburban)
Language UPC
Platform Cray X1, 1-8 processors)
33
Harder: Parallelizing Vector GIS
•(1/30) second Response time constraint on Range Query
• Parallel processing necessary since best sequential computer cannot meet requirement
• Blue rectangle = a range query, Polygon colors shows processor assignment
Set of Polygons
Set of Polygons
Display Graphics Engine
Local Terrain
Database
Remote Terrain
Databases
30 Hz. View Graphics
2Hz.
8Km X 8Km Bounding Box
High Performance
GIS Component
25 Km X 25 Km
Bounding Box
34
Data-Partitioning Approach
• Initial Static Partitioning
• Run-Time dynamic load-balancing (DLB)
• Platforms: Cray T3D (Distributed), SGI Challenge (Shared Memory)
35
DLB Pool-Size Choice is Challenging!
36
Hardest – Location Prediction
Nest locations Distance to open water
Vegetation durability Water depth
37
Ex. 3: Hardest to Parallelize
Name Model
Classical Linear Regression
Spatial Auto-Regression
εxβy
εxβWyy ρ
framework spatialover matrix odneighborho -by- :
parameter n)correlatio-(auto regression-auto spatial the:
nnW
• Maximum Likelihood Estimation
• Need cloud computing to scale up to large spatial dataset.
• However,
• Map reduce is too slow for iterative computations!
• computing determinant of large matrix is an open problem!
SSEnn
L2
)ln(
2
)2ln(ln)ln(
2
WI
38
Parallelizing Spatial Big Data on Cloud Computing
• Case 1: Compute Spatial-Autocorrelation Simpler to Parallelize
– Map-reduce is okay
– Should it provide spatial de-clustering services?
– Can query-compiler generate map-reduce parallel code?
• Case 2: Harder : Parallelize Range Query on Polygon Maps
– Need dynamic load balancing beyond map-reduce
– MPI or OpenMP is better!
• Case 3: Estimate Spatial Auto-Regression Parameters, Routing
– Map-reduce is inefficient for iterative computations due to expensive “reduce”!
– MPI, OpenMP, Pregel or Spatial Hadoop is essential!
– Ex. Golden section search, Determinant of large matrix
– Ex. Eco-routing algorithms, Evacuation route planning
39
Outline
• Motivation
• SBD Definitions & Examples
• SBD Infrastructure
• SBD Analytics
• Conclusions
40
Data Mining to Spatial Data Mining to SBD Analytics
• 1990s: Data Mining
• Scale up traditional models (e.g., Regression) to large relational databases (460 Tbytes)
• New pattern families: Associations : Which items are bought together? (Ex. Diaper, beer)
• Spatial customers
• Walmart: Which items are bought just before/after events, e.g. hurricanes?
• Where is a pattern (e.g., (diaper-beer) prevalent?
• Global climate change: tele-connections
• But faced challenges
• Independent Identical Distribution assumption not reasonable for spatial data
• Transactions, i.e. disjoint partitioning of data, not natural for continuous space
• This led to Spatial Data Mining (last decade)
• SBD raise new questions • May SBD address open questions, e.g. estimate spatial neighborhood (e.g., W matrix)?
• Does SBD facilitate better spatial models, e.g., place based ensembles beyond GWR?
• (When) Does bigger spatial data lead to simpler models, e.g. database as a model ?
• On-line Spatio-temporal Data Analytics
41
Spatial Data Mining Example 1: Spatial Prediction
Nest locations Distance to open water
Vegetation durability Water depth
42
Spatial Autocorrelation (SA)
• First Law of Geography
– “All things are related, but nearby things are more related than distant things. [Tobler, 1970]”
• Autocorrelation
– Traditional i.i.d. assumption is not valid
– Measures: K-function, Moran’s I, Variogram, …
Pixel property with independent identical
Distribution (i.i.d)
Vegetation Durability with SA
43
Parameter Estimation for Spatial Auto-regression
Name Model
Classical Linear Regression
Spatial Auto-Regression
εxβy
εxβWyy ρ
• Maximum Likelihood Estimation
• Computationally Expensive
• Determinant of a large matrix
• Iterative Computation
• Golden Section Search for
SSEnn
L2
)ln(
2
)2ln(ln)ln(
2
WI
framework spatial over matrix od neighborho - by - : n n W
parameter n) correlatio - (auto regression - auto spatial the :
44
Spatial Data Mining Example 2: Clustering
• Clustering: Find groups of tuples
• Statistical Significance
– Complete spatial randomness, cluster, and de-cluster
Inputs:
Complete Spatial Random (CSR),
Cluster,
Decluster
Classical Clustering
(K-means always finds clusters)
Spatial Clustering begs to differ!
Leftmost: Clusters not significant
Rightmost – declustered!
45
Spatial Data Mining Example 3: Spatial Outliers
• Spatial Outliers
– Locations different than neighbors
– Ex. Sensor 9
– Source: Traffic Data Cities
• Spatial Join Based Tests
46
Association Patterns
• Association rule e.g. (Diaper in T => Beer in T)
– Support: probability (Diaper and Beer in T) = 2/5
– Confidence: probability (Beer in T | Diaper in T) = 2/2
• Algorithm Apriori [Agarwal, Srikant, VLDB94]
– Support based pruning using monotonicity
• Note: Transaction is a core concept!
Transaction Items Bought
1 {socks, , milk, , beef, egg, …}
2 {pillow, , toothbrush, ice-cream, muffin, …}
3 { , , pacifier, formula, blanket, …}
… …
n {battery, juice, beef, egg, chicken, …}
47
Spatial Data Mining Example 4: Co-locations/Co-occurrence
• Given: A collection of
different types of spatial
events
• Find: Co-located
subsets of event types
• Challenge:
No Transactions
• New Approaches
– Spatial Join Based
– One join per
candidate is
Computationally
expensive!
48
SBD Opportunities 1: Estimate Spatial Neighbor Relationship
• SDM Limitation 1: Neighbor relationship is End-users’ burden !
– Colocation mining, hotspot detection, spatial outlier detection, …
– Example: W matrix in spatial auto-regression
– Reason: W quadratic in number of location
– Reliable estimation of W needs very large number data samples
• SBD Opportunity 1: Post-Markov Assumption
– SBD may be large enough to provide reliable estimate of W
– This will relieve user burden and may improve model accuracy
– One may not have assume
• Limited interaction length, e.g. Markov assumption
• Spatially invariant neighbor relationships, e.g., 8-neighbor
• Tele-connections are derived from short-distance relationships
Name Model
Classical Linear Regression
Spatial Auto-Regression
εxβy
εxβWyy ρ
49
SBD Opportunity 2: Place Based Ensemble of Models
• SDM Limitation 2: Modeling of Spatial Heterogeneity is rare
– Spatial Heterogeneity: No two places on Earth are identical
– Yet, Astro-Physics tradition focused on place-independent models
– Was it due to paucity of data ?
– Exception: Geographically Weighted Regression or GWR [ Fortheringham et al. ]
– GWR provides an ensemble of linear regression models, one per place of interest
• Opportunity 2: SBD may support Place based ensemble of models beyond GWR
– Example: Place based ensemble of Decision Trees for Land-cover Classification
– Example: Place based ensemble of Spatial Auto-Regression Models
– Computational Challenge:
• Naïve approach may run a learning algorithm for each place.
• Is it possible to reduce computation cost by exploiting spatial auto-correlation ?
50
SBD Opportunity 3: Bigger the SBD, Simpler the Model
• SDM Limitation 3: Complexity of Spatial Models
– Spatial models are usually computationally more expensive than traditional models
– Example: Spatial Auto-regression vs. Linear Regression
– Example: Geographically Weighted Regression vs. Regression
– Example: Colocation Pattern Mining vs. Association Rule Mining
– Confidence Estimation adds more costs, e.g. M.C.M.C. simulations
• SBD Opportunity 3: Bigger the SBD, Simpler the spatial models
– Sometime the complexity is due to paucity of data at individual places
• It forces one to leverage data at nearby places via spatial auto-correlation and spatial join!
– SBD may provide plenty of data at each place!
• This may allow place-based divide and conquer
• Build one model per place using local data and simpler model
– Challenges:
• Compare place-based ensembles of simpler models with current spatial models
• When does bigger data lead to simpler models?
• What is SBD from analytics perspective ?
– e.g., ratio of samples to number of parameters
51
SBD Opportunity 4:On-line Spatio-temporal Data Analytics
• SDM Limitation 3: Off-line Batch Processing
– Spatial models are usually learned in off-line batch manner
• Example: Spatial Auto-regression, Colocation Pattern Mining, Hotspot Detection
– However, SBD include streaming data such as event reports, sensor measurements
– SBD use cases include monitoring and surveillance requiring On-Line algorithms
• Example: Timely Detection of Outbreak of disease, crime, unrest, adverse events
• Example: Displacement or spread of a hotspot to neighboring geographies
• Example: Abrupt or rapid change detection in land-cover, forest-fire, etc. for quick response
• SBD Opportunity 4: On-line Spatio-temporal Data Analytics
– Purely local models may leverage time-series data analytics models
– Regional and global models are harder
• Spatial interactions (e.g., colocation, tele-connections) with time-lags
• Q? Can these be computed exactly in on-line manner?
• Q? If not, what are on-line approximations ?
52
Spatial Big Data (SBD)
• SBD Definitions
• SBD Applications
• SBD Infrastructure
• SBD Analytics
• Conclusions
53
Spatial Big Data (SBD)
• SBD are important to society
– Ex. Eco-routing, Public Safety & Security, Understanding Climate Change
• SBD exceed capacity of current computing systems
• DBMS Challenges
– Eco-Routing: Lagrangian frame, Non-Stationary Ranking
– Privacy vs. Utility Trade-offs
• Data Analytics Opportunities
– Post Markov Assumption – Estimate Neighbor Relationship from SBD
– Place based Ensemble Models to address spatial heterogeneity
– Bigger the spatial data, simpler may be the spatial models
– Online Spatial Data Analytics
• Platform Challenges
– Map-reduce – expensive reduce not suitable for iterative computations
– Load balancing is harder for maps with polygons and line-strings
– Spatial Hadoop ?
54
Acknowledgments
• HPC Resources, Research Grants
– Army High Performance Computing Research Center-AHPCRC
– Minnesota Supercomputing Institute - MSI
• Spatial Database Group Members
– Mete Celik, Sanjay Chawla, Vijay Gandhi, Betsy George, James Kang,
Baris M. Kazar, QingSong Lu, Sangho Kim, Sivakumar Ravada
• USDOD
– Douglas Chubb, Greg Turner, Dale Shires, Jim Shine, Jim Rodgers
– Richard Welsh (NCS, AHPCRC), Greg Smith
• Academic Colleagues
– Vipin Kumar
– Kelley Pace, James LeSage
– Junchang Ju, Eric D. Kolaczyk, Sucharita Gopal
